2.1. Political participation and communication on social media

Social media has fundamentally changed the way people participate in political discourse. By providing an online platform for individuals to engage in political discussions and express their political views, social media enables political communication globally, at scale, and in rapid succession. Social media provides new opportunities not only for ordinary citizens but also for journalists and politicians. Journalistic YouTubers, for example, aim to push journalism toward younger target groups, establish entertaining presentation styles, and strongly focus on audience interactions [16]. Politicians use social media for election campaigning [17] and to share their messages across different target groups [18]. However, in the remainder of this paper, we focus on political communication from the perspective of citizens and potential voters, as this also reflects the audience of influencers. Among this audience, political participation comes in various forms and shapes, including, for instance, liking and sharing political content and engaging in political discussions. Socio-demographics and other factors, in turn, affect the kind of people who are more likely to engage in political discourses on social media. For example, young people, in particular, are more likely to use social media as it is also their primary source of news and information [19].

On the one hand, these possibilities for online political communication go hand in hand with the promise of lower-threshold opportunities and the associated increase in political equality. Some studies have shown that political participation on social media can have a significant impact on political outcomes, particularly in terms of voter turnout and political efficacy [20,21]. However, access to social media by itself does not necessarily lead to an increase in political involvement, as a large part of the increased participation is due to the self-selection processes of politically active citizens. As other studies suggest, it is primarily people who have already been active offline who are participating in the political discourse online [22,23]. A key question, therefore, is how to engage citizens who are active on social media but not for political reasons. According to a study in Germany, reaching younger voters requires authentic and personal content, explaining politics in simple terms, and adjusting to communication cultures and specifics of online platforms [24]. Social media influencers excel at these requirements, bridging the connection between online political communication and authentic opinion leaders.

2.2. Influencers as authentic opinion leaders

While there are many definitions of influencers, especially in the marketing literature, a dominant definition across scientific fields has yet to be established. Abidin, for example, defines influencers as

“everyday, ordinary Internet users who accumulate a relatively large following on blogs and social media through the textual and visual narration of their personal lives and lifestyles, engage with their following in digital and physical spaces, and monetize their following by integrating ‘advertorials’ into their blog or social media posts. [...] It is a functional attribution that organizations apply to social media users such as bloggers, YouTubers, Instagrammers, etc., that are ascribed the ability to influence the organization’s stakeholders and thus become relevant to the organization.” [25]

On the one hand, it is clear from this definition that influencers present themselves as “ordinary people,” thus, a distinction is made between them and celebrities, such as film stars or well-known athletes [26]. On the other hand, influencers are characterized by the fact that they incorporate product placements or advertisements into their everyday narrative and thus monetize their profile. As ordinary people, influencers report on their everyday lives and attract viewers to their profiles through the (supposedly) authentic portrayal of their lives. It is precisely this authenticity that influencers embody that is the core of their recipe for success [27,28]. The impression of authenticity conveyed is presented in research as a decisive success factor [29]. However, many so-called “micro-influencers” with a comparatively small number of followers do not operate their profile full-time and see their income from possible collaborations as a sideline or hobby.

Influencers are frequently characterized by their authenticity and interactivity, leading to ongoing debates about their role as authentic opinion leaders. [3,27,28]. In our work, we follow the established concept of Paul Lazarsfeld and consider opinion leaders as individuals who are particularly influential in shaping the opinions and behaviors of others within their social networks. Certain knowledge and expertise are attributed to them, which enables opinion leaders to communicate their opinions and ideas to others effectively [30]. In the 1940s, when the concept gained popularity in the academic debate, Lazarsfeld and others suggested that these individuals play an important role in shaping public opinion and behavior through two-step-communication flow. The two-step flow of communication theory posits that media messages influence the public indirectly, primarily through opinion leaders who first interpret and then relay these messages to others [31]. Although the concept of opinion leadership is not new, through social media and the associated communication options, it gains new relevance. While the original concept refers to face-to-face interactions, i.e., conversations with friends and acquaintances, digital opinion leaders differ in that they interact with their followers online. Unlike traditional opinion leaders, influencers can reach a much larger audience due to algorithm-driven visibility and network effects, yet they still cultivate perceptions of authenticity and peer-like relationships through parasocial interactions [27,32]. Potential opinion leaders can now build a greater network and reach a higher audience (respectively, higher influence). Recent research also suggests that the dissemination of information through social media is evolving into a multi-step flow with different kinds of opinion leaders, adopters, and influencers [33].

However, opinion leadership in isolation is insufficient for connecting political communication and influencers to develop a unified definition of political influencers. Within the limited but growing share of studies on political influencers, quite different concept definitions can be found [11,13]. To give one example from the introduction of a special issue on the topic: “We define political influencers as content creators that endorse a political position, social cause, or candidate through media that they produce and/or share on a given social media platform” [7]. The authors further outline that this definition can also include politicians and journalists. Similarly, Bause speaks of political social media influencers. She also refers to the concept of opinion leadership in her definition and describes influencers as “as self-created personal brands” [27]. Depending on the focus of a study, this rather broad definition of political influencers may be an excellent starting point. For other studies, particularly those with a strong connection to the concept of influencers as ordinary citizens establishing themselves as authentic opinion leaders, a more restrictive conceptualization of political influencers may be necessary [33]. Such conceptualizations may, for instance, exclude established elites such as party politicians or popular journalists as entities that should be included in an appropriate sampling frame for empirically analyzing the content of political influencers. For our case study, we refer to Abidin’s definition and focus on ordinary users rather than journalists and politicians who share political content. This narrower conceptualization allows us to specifically analyze influencers as a distinct phenomenon characterized by their perceived authenticity, interactivity, and grassroots-level impact.

Platform

One of the first decisions when studying influencers is what online platforms should be selected for data collection. This decision has significant implications, as social media platforms are tailored toward different target audiences, with varying user interfaces and varying models for maximizing engagement to accompany the underlying business models, which are predominantly based on advertisement revenue. At the same time, these affordances lead to some platforms being more popular among influencers than others. Besides decisions based on substantive criteria, accessing platform data is an additional technical hurdle challenging scholars who wish to study social media. While in principle, data from all platforms should be accessible to researchers via endeavors such as the Digital Service Act [34], in practice, data access is becoming more difficult. The situation about which application programming interfaces (APIs) for retrieving platform data are available for scientific use and at what cost is rapidly evolving. In the Post-API Age [35], most social media platforms do not provide means to access their content programmatically. This may lead to scholars applying data scraping procedures and violating terms of service, posing additional legal and ethical challenges [36].

Audience

After identifying the platforms of interest, scholars must decide upon the audience political influencers need to have accumulated to be considered for further analysis. A simple categorization approach is based on the number of followers or subscribers, as it offers insights into the account’s reach. It indicates how many people are regularly exposed to its content. Profiles are often categorized into different tiers based on follower counts. It is important to note that while the number of followers has important consequences regarding the potential reach, influencers with smaller audiences can still achieve the status of microcelebrities [14,37] and use corresponding techniques to increase their credibility. Authenticity and credibility can be at least or even more important for political influencers than many followers [38]. Depending on the research question, it may also be relevant to consider the engagement rate as a selection criterion beyond the follower count. Engagement rate is a metric that measures the level of interaction and engagement users have with social media content. It considers actions like likes, comments, shares, and clicks, providing a more comprehensive view of user involvement beyond just likes [39]. This metric is valuable because it reflects engagement quality, audience sentiment, potential virality, and content effectiveness. By analyzing engagement rates, researchers can gain insights into user behavior and assess the impact of their social media strategies.

Account type

While metrics such as the number of followers provide initial insights into the reach and sphere of visibility of influencers, many research questions require a deeper examination of their profiles. Therefore, the next step discussed focuses on the account type. While content creators prioritize self-produced content, which real individuals typically portray, other accounts are predominantly used to disseminate memes [40]. Another possibility is using accounts by multiple individuals, such as a couple of accounts or those representing organizations or events. Additionally, the rise of virtual characters, also known as AI influencers, is gaining momentum. AI influencers are virtual personas that engage with audiences on social media platforms, often mimicking human-like behavior and characteristics [41]. For research questions that primarily revolve around the content of a post or account, this distinction may be of secondary importance. However, when the focus is set on influencers as authentic human opinion leaders, this differentiation becomes significant and should be considered when selecting cases.

Profile description

As mentioned above, there are many definitions of influencers that are shared across different areas. While some subsume politicians, journalists, or celebrities under the term influencer, some intentionally focus on ordinary internet users, bloggers, and content creators. Fischer et al., for example, follow the second line by only analyzing influencers without an “institutional background in the established media and/or political system” [42]. De Gregorio includes politicians and content creators but differentiates between three types: Politicians who present themselves as Social Media Influencers, Social Media Influencers who appear as political opinion leaders and Social Media Influencers turned politicians [43]. Depending on the conceptualization of political influencers, scholars may use external resources to identify interest profiles. For example, official social media accounts of public figures are often available in online repositories or curated datasets of past research. In any case, qualitative examinations of social media profile descriptions in conjunction with available content help identify political influencers and exclude entities not part of the desired research sample.

Sampling

Once a sufficient set of influencer accounts is identified, scholars must decide upon an appropriate sampling strategy for capturing the content of interest. One of the biggest challenges when studying political influencers is that the target population of interest is in constant flux, making it hard to analyze a stable sample over time. Therefore, a number of related decisions have to be considered. For one, an important question is what kind of content should be sampled over what time period. Some content types, such as Instagram stories (see next paragraph), are usually only available for a limited time. Moreover, some content, such as hateful or xenophobic posts by right-wing influencers, may frequently be deleted from online platforms. Both cases require scholars to adjust their sampling frequency accordingly. Samples might also be adjusted in case of prolonged observation periods as some accounts of political influencers vanish, while new accounts matching the selection criteria may appear.

Content

The content that is distributed on social media platforms varies greatly depending on the platform and offers users the opportunity to express themselves via text, images, and videos. The platforms are subject to constant change regarding which formats currently work best for reaching large audiences. One example is Instagram, which initially focused on images, leading to influencers mainly producing content favored by the platform’s algorithm [44]. Studies also show that images with faces achieve a higher reach than other image types [45]. In the meantime, however, Instagram has evolved and is increasingly promoting so-called reels, short videos that are given preferential treatment in the algorithm. The situation is similar on TikTok, with a strong focus on videos. For researchers working with social media content, this poses the challenge of working with different modalities (image, video, text, sound) and the constant need to follow trends about how content is produced and distributed by influencers. Content also includes possible interactions between audiences and influencers, for instance, in the form of user comments. It may further be relevant for researchers to distinguish between self-generated content and re-posts of other accounts. Capturing the content of interest lastly includes technical considerations regarding data retrieval and data storage, as handling and processing text, image, and video data can be a challenging endeavor.

Advertisement

With advertisement being the primary source of revenue for most social media platforms and also for influencers, its role in the content produced by political influencers deserves special attention. Traditional influencers, as well as political influencers, can seamlessly integrate a mixture of advertisements and topical content into their social media feeds [3,5]. Such advertisements can come in many forms, such as brand hashtags, unpaid or paid product features, and official partnerships, but many (paid) product recommendations are not officially labeled as such. For political influencers, in particular, advertisements can conflict with authenticity and credibility. When political influencers promote sustainable products or speak out in favor of supporting political parties while receiving monetary compensation at the same time, followers may question whether the content is made up of authentic opinions or merely paid propaganda.

In summary, we highlighted a couple of key factors to be considered for the conceptualization and operationalization of political influencers and the collection of their generated content. As indicated in Fig 1, this is an iterative process in practice, requiring multiple adjustments before finalizing a pipeline for data collection. In what follows, we proceed with a case study demonstrating a quantitative approach to studying political influencers and gaining new insights into their relevance for elections.

Content produced by political influencers

One of the main questions of interest for our case study is whether the dissemination of political posts makes up for the majority of their produced content or is a rather niche phenomenon. While political topics, such as sustainability and environmental protection, can be discussed in isolation, they often reference political actors or events. Especially when it comes to elections, another key question is whether political influencers support or disapprove political entities in the content they produce. At last, we also aim to quantify how much advertisement can be observed in the posts of political influencers, leading to research questions RQ1-RQ3:

1. RQ1: How much political content is produced by political influencers?
2. RQ2: To what extent do political influencers support or disapprove political actors or events?
3. RQ3: How much advertisement is included in the posts of political influencers?

Following the scheme outlined in Fig 1, we first decided on the platform. We focused on Instagram, as it was arguably the most prevalent platform for content creation of political influencers in Germany during this period [46]. We furthermore use a combination of keyword searches (including terms and hashtags with political connotations), network properties (e.g., searching through followees and followers of relevant accounts), and in-depth, qualitative inspections of profiles to identify influencer accounts. Second, for selecting influencer profiles for our sample, we considered the follower count and set our minimum threshold for inclusion at least 10,000 followers. This represents so-called micro influencers [51] but also includes other influencer accounts with higher follower counts. In a third step, we further preselected through a look at the account type. Indications of the account type are, for example, the self-description of the profiles (bio) and categories defined by Instagram, which can, however, also be set up by the profile owners. Fourth, as we are mainly interested in “original” influencers and not just influential accounts, we do not consider politicians or journalists. Rather, we focus on ordinary internet users sharing self-produced details of their lives & subsequently becoming influential by accumulating a large follower base [25], who produce at least some content related to politics. In the final step, we examine whether profiles have uploaded content with paid partnerships or advertising. Our final sample includes 98 accounts of political influencers who match our criteria. After identifying our main sample of interest, we used the Crowdtangle API [52] combined with additional data scraping procedures to capture the content produced by all accounts. We first use Crowdtangle API daily to get information about all posts produced by influencers. In the second step, we scrape videos, images, and photos of all posts. The data was collected between June 2021 and September 2021. We extracted a total of 3.164 posts, which were liked by around 30 million people and received around 335,000 comments at the end of our data collection period. Our data not only allows us to provide a one-time categorization of influencer content but also to identify time trends and evaluate influencers to adjust their posts in light of the federal elections.

However, a significant challenge is that our concepts of interest are notoriously difficult to measure. This is due to several reasons: First, the Instagram content is multimodal, consisting of text, photos, and videos. Despite several advances in machine learning and artificial intelligence for multimodal data [53], it is still difficult to identify multiple concepts across data types. Second, our concepts of interest cannot be captured via simple procedures. For one, there is no straightforward mapping from a definition to the operationalization of what can be considered political content - particularly, in a broad sense, covering a vast range of topics.

Moreover, while Instagram has integrated platform features for declaring paid sponsorships, these features are used only very infrequently. In combination, these challenges restrict methods for fully automated concept detection. For this reason, we employ crowdsourcing for human-assisted coding, a concept that has been applied successfully for coding political data in the past [54]. We use the platform Prolific to recruit a sample of German-speaking raters with a balanced gender ratio. Regarding the quality of data obtained via crowdsourcing, recent work has highlighted that a substantial share of crowd workers use Large Language Models for task completion [55]. While we can never fully rule out biases in our data, this concern is very unlikely to affect our task, as models were not nearly as proficient nor as easy to use as recent models at the time of data collection in 2021. This applies particularly to working with image data, a crucial component of our task. In addition, Prolific has been found to provide data with higher quality than other providers and is explicitly designed for scientific research [56]. For our task, we recruited a total of 357 participants and also paid above-average rates compared to other tasks on the platform with similar time investment.

Using the R-Shiny framework [57], we created a web application for the coding procedure. For each coder, a random subset of posts was shown. In addition to text, we showed up to four images for multi-image posts and the video thumbnail for video posts. For each post, coders were asked to indicate whether it includes political content (RQ1), support or disapproval of political actors (RQ2), and/or advertisement (RQ5). Our related coding instructions, as well as several example posts covering each of the categories, are available in the supplementary materials S1 and S2 in S1 Text. For each post, we accumulate at least three different coders, resulting in about 10,700 total ratings. In the last step, we identify the majority mode of codings for each category among coders and discard posts without modal agreement. We calculated these shares of posts without a modal agreement for each category: ~ 5% for support or disapproval of political entities, ~ 2% for advertisement, and ~ 1% for political content. In other words: we have a majority vote for 95% of posts related to support or disapproval of political actors, 98% related to advertisement, and 99% related to political content. These numbers suggest reasonable agreement between coders. Through qualitative examination, we found that the discarded posts were predominantly ambivalent with little meaning and thus not relevant to our case study. This procedure resulted in a final sample of 2.915 posts, 92% of our initial sample. At the time of data collection, these posts generated a cumulate reach of about 312 thousand comments and 29 million likes on Instagram.

Potential for affecting election outcomes

While the content produced by political influencers in the run-up to the election provides valuable insights into political influencers, analyzing it in isolation is only of limited use for evaluating the relevance of election outcomes. First, it is important to determine which kind of voters are likely to be in contact with influencers’ content in the first place. Second, consuming influencer content does not necessarily affect personal voting decisions. In addition, it is important to understand whether the content of influencers provided useful information for the decision-making of voters. This leads us to our research questions RQ4-RQ5:

1. RQ4: Which voters are following the content of influencers?
2. RQ5: To what extent do voters consider influencers’ content relevant to their election-related decision-making?

To answer these questions, we collaborated with the Online Access Panel provider Respondi to run a follow-up survey for the German federal elections. We restricted the participant pool to people eligible to vote and introduced sampling quotas by age, sex, education, and federal state according to the German census. In our supplementary material S4 in S1 Text, we compare the socio-demographic composition of our sample to the census. As expected, our sample composition regarding sex, age, education, and federal state, is well-aligned with the target population of German citizens. Our survey was conducted roughly three weeks after the election, with a field time from October 19 to October 25 and a total of 1.107 participants. Informed consent was acquired via the access panel provider. For our questionnaire, we mostly relied upon items from other large surveys, in particular, the German Longitudinal Election Study (GLES) [58], and added influencer-related questions. An overview of data quality checks is included in our supplementary material S5 in S1 Text. We compared voting shares from our survey with those from the GLES Post-Election Study and the actual results of the federal election. We find a high similarity between voting shares from our survey and those of both GLES and the election results. At last, we compared self-collected participant demographics to those provided by Respondi and removed a few speeders, resulting in a final sample of 1.059 survey participants. As speeders, we define all participants who either did not complete the questionnaire in full or took less than 120 seconds to complete it. Note that including or excluding speeders has no substantial effect on our results [59].